Fine-tuned model unable to answer prompts from training data